Port reusable audio preprocessing components from VoiceMemoApp

Context

[VoiceMemoApp](https://github.com/paulbauriegel/VoiceMemoApp) is an Android application for fully on-device voice recording and transcription.

It already contains several audio preprocessing components that overlap with functionality we want to provide as part of SKaiNET-audio.

The goal of this issue is to move the reusable parts into SKaiNET-audio so that applications such as VoiceMemoApp can use the library instead of maintaining their own DSP implementation.

Scope

Port / implement the following functionality based on the VoiceMemoApp implementation:

• MixedRadixFft

  • Radix 2 and 5
  • allocation-free processing
  • support for FFT sizes required by Whisper and Parakeet

• Polyphase sinc resampler with Blackman window

• Streaming WavAudioReader

  • PCM16
  • PCM24
  • PCM32
  • Float32
  • indexed/windowed access without loading the complete file into memory

• Whisper-compatible 128-Mel frontend

• MelConfig.whisperV3() preset

Where useful, the existing VoiceMemoApp implementation can serve as the reference implementation.

Validation

Add shared golden/reference tests for the 128-Mel frontend, ideally generated from the corresponding Hugging Face / Python preprocessing pipeline.

Suggested test data:

• 2–3 audio clips
• at least one clip shorter than 30 seconds
• comparison against the reference Mel output with an appropriate numerical tolerance

This should give us a reproducible reference that can also be used by VoiceMemoApp.

Integration goal

Once the functionality is available in SKaiNET-audio, create a follow-up PR for VoiceMemoApp that replaces the corresponding local DSP implementation with the SKaiNET-audio dependency.

This gives us a real-world integration test and should remove a significant amount of duplicated audio preprocessing code from the application.

Open points

• Confirm licensing / attribution for code derived from VoiceMemoApp
• Confirm FFT/window configuration for Parakeet
• Implement MixedRadixFft
• Replace direct DFT in MelSpectrogram
• Add polyphase sinc resampler
• Extend WAV reader
• Add MelConfig.whisperV3()
• Add 128-Mel golden tests
• Validate against VoiceMemoApp
• Follow-up PR to migrate VoiceMemoApp to SKaiNET-audio

Reference

VoiceMemoApp

• VoiceMemoApp: https://github.com/paulbauriegel/VoiceMemoApp